Compute the current learning rate under exponential or step decay schedules, for scheduler tuning and training-curve diagnosis.
指数衰减常用于后期精调。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| Exponential decay | the learning rate is multiplied by a fixed factor each epoch or step | PyTorch lr_scheduler docs |
| Cosine annealing | the learning rate follows a cosine curve from the initial value down to a floor | Loshchilov & Hutter 2017, SGDR |
| Effect | shrinking the step size helps the optimiser settle into a lower loss | Robbins & Monro 1951, Ann. Math. Stat. |